Home Available Analyses FAQ Docs
Sign In

Emergent Property Inference (Bittner et al. 2019)


Analysis description

EPI brings recent machine learning techniques – the use of deep generative models for probabilistic inference – to bear on the problem of learning distributions of parameters that produce the specified properties of computation in theoretical models. Importantly, the techniques introduced offer a principled means to understand the implications of model parameter choices on computational properties of interest. This implementation of EPI performs hyperparameter optimization for a 2D LDS model constrained to produce a periodic signal of given frequency. Used to demo NeuroCAAS and EPI at COSYNE 2020.

Useful links
Emergent Property Inference (Bittner et al. 2019) Paper Link
Emergent Property Inference (Bittner et al. 2019) Github Repo Link
Emergent Property Inference (Bittner et al. 2019) Bash Script Link
Emergent Property Inference (Bittner et al. 2019) Demo Link
How to use this analysis

This analysis trains an EPI model to produce a periodic signal, starting at different hyperparameter values.
Args:
  -Input: (json file) A JSON file formatted with a single value: the random seed used to initialize hyperparameters, like so:
    {
        "random_seed": INT
    }
  -Config: (yaml file) A YAML file specifying NeuroCAAS parameters for the job duration expected. No analysis specific parameters required. Click on the config provided for an example.

Returns:
  -Folder of results providing trained model, optimization history, and diagnostic output. See "epi_opt.mp4" for a video of learned parameter distributions.


You must login to use an analysis.

  • Twitter
  • Github
  • Dribbble